Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards Robust Recommendation via Decision Boundary-aware Graph Contrastive Learning

Jiakai Tang, Sunhao Dai, Zexu Sun, Xu Chen, Jun Xu, Wenhui Yu, Lantao Hu, Peng Jiang, Han Li

TL;DR

This work tackles robustness and data sparsity in graph-based recommender systems by introducing RGCL, a decision boundary-aware graph contrastive learning framework. It jointly optimizes semantic-preserving, hard contrastive views and margin-maximizing adversarial perturbations, guided by global user-user and item-item relations. The authors provide theoretical analysis of hardness-aware learning and robustness, and demonstrate consistent, statistically significant gains across five public datasets over twelve baselines, along with strong ablations and robustness studies. The approach yields faster convergence and improved performance, especially for sparse users and long-tail items, making it practically impactful for real-world recommendation systems.

Abstract

In recent years, graph contrastive learning (GCL) has received increasing attention in recommender systems due to its effectiveness in reducing bias caused by data sparsity. However, most existing GCL models rely on heuristic approaches and usually assume entity independence when constructing contrastive views. We argue that these methods struggle to strike a balance between semantic invariance and view hardness across the dynamic training process, both of which are critical factors in graph contrastive learning. To address the above issues, we propose a novel GCL-based recommendation framework RGCL, which effectively maintains the semantic invariance of contrastive pairs and dynamically adapts as the model capability evolves through the training process. Specifically, RGCL first introduces decision boundary-aware adversarial perturbations to constrain the exploration space of contrastive augmented views, avoiding the decrease of task-specific information. Furthermore, to incorporate global user-user and item-item collaboration relationships for guiding on the generation of hard contrastive views, we propose an adversarial-contrastive learning objective to construct a relation-aware view-generator. Besides, considering that unsupervised GCL could potentially narrower margins between data points and the decision boundary, resulting in decreased model robustness, we introduce the adversarial examples based on maximum perturbations to achieve margin maximization. We also provide theoretical analyses on the effectiveness of our designs. Through extensive experiments on five public datasets, we demonstrate the superiority of RGCL compared against twelve baseline models.

Towards Robust Recommendation via Decision Boundary-aware Graph Contrastive Learning

TL;DR

This work tackles robustness and data sparsity in graph-based recommender systems by introducing RGCL, a decision boundary-aware graph contrastive learning framework. It jointly optimizes semantic-preserving, hard contrastive views and margin-maximizing adversarial perturbations, guided by global user-user and item-item relations. The authors provide theoretical analysis of hardness-aware learning and robustness, and demonstrate consistent, statistically significant gains across five public datasets over twelve baselines, along with strong ablations and robustness studies. The approach yields faster convergence and improved performance, especially for sparse users and long-tail items, making it practically impactful for real-world recommendation systems.

Abstract

In recent years, graph contrastive learning (GCL) has received increasing attention in recommender systems due to its effectiveness in reducing bias caused by data sparsity. However, most existing GCL models rely on heuristic approaches and usually assume entity independence when constructing contrastive views. We argue that these methods struggle to strike a balance between semantic invariance and view hardness across the dynamic training process, both of which are critical factors in graph contrastive learning. To address the above issues, we propose a novel GCL-based recommendation framework RGCL, which effectively maintains the semantic invariance of contrastive pairs and dynamically adapts as the model capability evolves through the training process. Specifically, RGCL first introduces decision boundary-aware adversarial perturbations to constrain the exploration space of contrastive augmented views, avoiding the decrease of task-specific information. Furthermore, to incorporate global user-user and item-item collaboration relationships for guiding on the generation of hard contrastive views, we propose an adversarial-contrastive learning objective to construct a relation-aware view-generator. Besides, considering that unsupervised GCL could potentially narrower margins between data points and the decision boundary, resulting in decreased model robustness, we introduce the adversarial examples based on maximum perturbations to achieve margin maximization. We also provide theoretical analyses on the effectiveness of our designs. Through extensive experiments on five public datasets, we demonstrate the superiority of RGCL compared against twelve baseline models.
Paper Structure (43 sections, 1 theorem, 24 equations, 9 figures, 4 tables, 1 algorithm)

This paper contains 43 sections, 1 theorem, 24 equations, 9 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. GNN-based Recommendation
  4. GCL-based Recommenders
  5. Adversarial Robustness
  6. Our Approach: RGCL
  7. Overall Framework
  8. Decision Boundary-aware Perturbation
  9. Relation-aware Contrastive Learning with Perturbation Constraints
  10. Perturbation-constrained Contrastive Augmentation
  11. Relation-aware Adversarial-Contrastive Augmentation
  12. Multi-View Contrastive Learning
  13. Towards Margin Maximization via Adversarial Optimization
  14. Model Training
  15. Joint Optimization Objective
  16. ...and 28 more sections

Key Result

Theorem 1

Gradient descent on $\psi(g(x+\mathbf{\Delta}^*;\bm{\theta}))$w.r.t.$\bm{\theta}$ with a proper step size increases $d(x;\bm{\theta})$, where $\mathbf{\Delta}^*=\arg \max_{g(x+\mathbf{\Delta};\bm{\theta})> 0} \Vert \mathbf{\Delta} \Vert$ is the maximum perturbation given the current $\bm{\theta}$.

Figures (9)

  • Figure 1: An overview of two types of representative GCL-based recommenders. To facilitate the presentation, we only show a single user and item with injected noise. However, in practice, the semantic-aware GCL-based methods should integrate perturbations to all graph nodes.
  • Figure 2: Overall framework of our proposed dynamic decision boundary-aware graph contrastive learning framework RGCL.
  • Figure 3: Model convergence analysis w.r.t training epochs on the ML-1M and Yelp datasets.
  • Figure 4: Recommendation performances at different level of data sparsity and item popularity. The black dashed line represents no performance improvement or decline.
  • Figure 5: Hyper-parameter analysis w.r.t. $\alpha$, $L$, $\tau$. The top shows the experimental results on ML-1M and the bottom shows the results on Yelp.
  • ...and 4 more figures

Theorems & Definitions (1)

  • Theorem 1